Graph neural network

If you can represent your data as a graph, you can use a graph neural network to perform inference on it.
GNN operate on global graph embeddings or local embeddings in each node or edge in the graph.
Hence, a GNN allows you to output predictions on each node, each edge, or the entire graph.

Introduction

Structure

A graph neural network consists of layers which operate on graphs.
Typically, this means one or more GNN layers to get new embeddings along the graph.
Then a standard MLP can be used to parse each embedding into logits or values.

At each layer, you get the following features during inference:

• The graph structure
• Node embeddings
• Edge embeddings
• Graph embeddings

Message Passing Layer

A standard GNN layer consists of pooling functions followed by update functions.

Pooling

There are many types of pooling to choose from:

• For nodes, you can add the node embedding to connected node embedding or connected edge embeddings.
• Similarly for edges, you can add its embedding to connected node embeddings.
• For the entire graph, you can add all node embeddings together.

After pooling, you can use update function (e.g. MLP) to update the embeddings/state at each node, edge, and for the entire graph.

Resources

• A Gentle Introduction to Graph Neural Networks (2021) by Google Research
• Understanding Convolutions on Graphs (2021) by Google Research